Research map / Image generation / Editing and style transfer

CADMorph: Geometry‑Driven Parametric CAD Editing via a Plan–Generate–Verify Loop

NeurIPS 2025 · Image generation › Editing and style transfer · approach: Localized and attribute-specific editing · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

CADMorph edits a CAD parametric construction sequence to match a target geometry via an iterative plan-generate-verify loop using a diffusion P2S model's cross-attention, an LLM infiller (FlexCAD-style MPP), and latent-distance verification.

Problem

Geometry-driven parametric CAD editing: update a parametric construction sequence so its rendered shape matches a target shape while preserving structure, semantic validity, and shape fidelity, under scarce editing triplet data.

Approach

Use pretrained domain models at inference time: cross-attention maps of a parameter-to-shape latent diffusion model locate segments to mask, a masked-parameter-prediction LLM infills candidates, and the diffusion model's latent space scores candidates against the target shape, with a cross-iteration priority queue; no triplet training data is needed.

Key results

• CADMorph reaches IoU 0.687 and mean CD 0.009 versus 0.447 IoU and 0.029 CD for the best prior FlexCAD and 0.548 IoU for CAD-Diffuser.
• Invalid rate drops to 3.1% compared with 25.1% for GPT-4o and 5.7% for CAD-Diffuser.
• Ablation removing the verification stage drops IoU from 0.687 to 0.517; removing planning drops it to 0.447.

Measured effects reported in the paper

Surprising findings

• Removing the cross-iteration priority queue lowers IoU from 0.687 to 0.619, showing that retaining earlier candidates matters.
• Reasoning variants o4-mini and o4-mini-high do not improve over GPT-4o and sometimes do worse.

Limitations

• Inference latency from several plan-generate-verify iterations.
• Evaluated only on the CAD-Editor test set, whose models are simpler than real-world assemblies.
• Some failure cases where edited sequences fail to render valid shapes.

Experimental setup

Datasets: DeepCAD CAD-Editor test set (2k)

Benchmarks: CAD-Editor 2k test set

Metrics: IoU mean Chamfer Distance median Chamfer Distance JSD invalid rate edit distance human rank

Baselines: GPT-4o o4-mini o4-mini-high CAD-Diffuser FlexCAD

Builds on: SDFusion FlexCAD prompt-to-prompt LoRA

Closest papers by meaning

• CAD-Editor: A Locate-then-Infill Framework with Automated Training Data Synthesis for Text-Based CAD Editing (ICML 2025 · similarity 0.88)
• CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.87)
• ShapeWalk: Compositional Shape Editing Through Language-Guided Chains (CVPR 2024 · similarity 0.87)
• CAD-GPT: Synthesising CAD Construction Sequence with Spatial Reasoning-Enhanced Multimodal LLMs (AAAI 2025 · similarity 0.86)
• Shape of Thought: Progressive Object Assembly via Visual Chain-of-Thought (ICML 2026 · similarity 0.86)
• cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.86)